Background: Given that the classification of EEG signals and feature extraction from these signals play a fundamental role in BCI systems, providing a reliable and effective classification algorithm for improving BCI applications is highly desirable. Various methods have been proposed for decoding and classifying EEG signals. These methods usually extract patterns as a first step and then train a classifier to identify the user's intention. While these methods have achieved satisfactory performance, they usually assume that test data have a similar training/distribution process. However, in many BCI applications, this is often not the case. In fact, different psychological states during data acquisition and human equipment can lead to changes in EEG data. In addition, data from a specific motor imagery (such as right arm movement) may not be the same for different individuals, and in general, there may be a limited number of training samples available.
Objective: To investigate and address the challenges of existing brain-computer interface systems.
Methodology: In this study, a dataset consisting of signals from 109 subjects in 4 motor imagery and real movement states was used. In the preprocessing stage, the Butterworth filter and the fastICA algorithm were used to clean the data and remove artifacts. Then, 136 features from various domains, such as time domain, frequency domain, and transform domain, were extracted from each EEG section, and using the mrmr algorithm, an optimal set of features was selected as input to the classifier. Next, the extracted features from various EEG channels were considered as a spatial series, and this spatial series was used as input to the LSTM classifier.
Findings: The evaluation results indicate that the proposed algorithm can achieve an accuracy of over 88% in person-to-person assessment and an accuracy of over 70% on average across different individuals. Additionally, it can be said that frequency-domain features have more imp